NASA and IBM have developed an open-source Lunar Foundation Model designed to help scientists analyse Moon data more efficiently. Trained primarily on about 2 million image tiles from NASA’s Lunar Reconnaissance Orbiter, the model incorporates additional data from missions including GRAIL, Lunar Prospector and Japan’s SELENE. Researchers can use the system to identify impact craters, study volcanic formations and assess potentially ice-rich terrain near the lunar poles. The model is publicly available through Hugging Face and its codebase through GitHub.
NASA-IBM Lunar Foundation Model Launch Marks New Era for AI-Powered Lunar Science
NASA and IBM say they will launch an open-source artificial intelligence model designed for lunar research that will allow scientists around the globe to analyse Moon data in less time and support future exploration missions.
The NASA-IBM Lunar Foundation Model will revolutionise lunar science by enabling scientists to analyse vast amounts of Moon data more efficiently. The model was developed in a collaboration between NASA, IBM Research, and academia and is among the first open-source artificial intelligence systems designed specifically for lunar exploration.
The model was mostly trained on data collected by NASA’s Lunar Reconnaissance Orbiter (LRO) and is hosted publicly on Hugging Face, with its entire codebase available via GitHub. The effort is designed to allow scientists everywhere to use the most advanced lunar analysis tools.
The model allows scientists to evaluate the rugged surface of the Moon, study its geologic history, and plan for future science missions, NASA says. The system is designed to accelerate analysis and eliminate extensive manual data processing.
NASA has spent decades building an extensive scientific record of the Moon, said Kevin Murphy, NASA’s Chief Science Data Officer and Acting Chief Data and AI Officer at NASA Headquarters in Washington.
Collecting information is just part of NASA’s mission, Murphy said. Making data easier for scientists to explore and use is just as important. The NASA-IBM Lunar Foundation Model shows how artificial intelligence can transform large scientific datasets into new discoveries, he said.
Lunar Reconnaissance Orbiter Data Powers AI Model
The foundation model was trained on about 2 million image tiles from NASA's Lunar Reconnaissance Orbiter mission. These included over one million high-resolution camera images at one-meter resolution and close to 964,000 multispectral images at 100-meter resolution.
The data set from the Lunar Reconnaissance Orbiter covers the majority of the moon’s surface and is larger than the total data set produced by all other NASA planetary missions combined, NASA said. The archive was huge and provided a solid base for training the artificial intelligence system.
The model’s knowledge of lunar terrain and composition was also expanded with additional training data from NASA’s Gravity Recovery and Interior Laboratory (GRAIL), NASA’s Lunar Prospector mission, and the Japan Aerospace Exploration Agency’s Selenological and Engineering Explorer mission.
Traditional machine-learning systems are designed for specific tasks and trained on specific datasets, while foundation models are pre-trained on massive datasets and then adapted to a wide range of scientific applications with minimal additional training. This approach improves efficiency while widening the possibilities of research.
The model enables researchers to find impact craters, identify strange volcanic formations, and determine regions of potentially ice-rich terrain at the lunar poles. These capabilities are deemed important for the understanding of lunar evolution and for supporting long-term exploration goals.
Mapping Lunar Ice and Geological History
One of the approach's most potent results was in estimating polar ice stability. NASA said the system performed as well as or better than several existing baseline models, while preserving fine-scale patterns associated with potential ice-rich areas.
Ice deposits in extremely cold, permanently shadowed areas near the lunar poles could have remained frozen for billions of years. Scientists believe that studying these regions could tell them about the history of the moon, and provide resources that would be useful for future missions.
The model also improves detection of irregular mare patches, unusual volcanic features that seem relatively young compared to established estimates of lunar geological cooling. Mapping these structures may provide clues to help researchers refine theories of the Moon’s thermal evolution.
Another important application is the detection of craters. This model permits an automated identification and measurement of impact craters, thus advancing the effort to date lunar surfaces and to reconstruct major events in the history of the solar system.
NASA demonstrated the technology with Lunar Reconnaissance Orbiter imagery before and after an impact from a SpaceX rocket body near the Einstein crater. The system correctly identified the existing craters and highlighted the newly created impact feature.
This project is part of a broader NASA AI for Science strategy led by the Office of the Chief Science Data Officer. This effort is part of a series of NASA-IBM partnerships that include the Prithvi Models for Earth observation and the Surya Model for space weather prediction, according to Sprouts News.
The effort included scientists from the Universities Space Research Association, SETI Institute, University of Maryland Baltimore County, Howard University, NASA Ames Research Centre, NASA Goddard Space Flight Centre, and NASA Headquarters. NASA and IBM are releasing datasets, benchmarks,s and tools through the open-source TerraTorch platform to foster global collaboration and advance the future of lunar exploration.
